3433 - A Deep Learning-Based Method for Predicting Out-of-Field Radiotherapy Dose for Pediatric Wilms Tumor Patients

M. Saha¹, J. W. Jung², M. Gopalakrishnan³, K. John⁴, C. Lee¹, and M. Mille¹; ¹National Cancer Institute, National Institutes of Health, Rockville, MD, ²Department of Radiation Oncology, Brody School of Medicine, East Carolina University, Greenville, NC, ³Northwestern Memorial Hospital, Chicago, IL, ⁴Department of Radiation Oncology, Northwestern University Feinberg School of Medicine, Chicago, IL

Purpose/Objective(s): Unintended out-of-field dose has been implicated as an important contributor to late health effects following radiotherapy. However, there are gaps in knowledge about organ dose tolerance. Furthermore, the out-of-field dose cannot be estimated accurately by the treatment planning system, but must instead be calculated by time-consuming Monte Carlo radiation transport simulation. Therefore, to support clinical treatment optimization and research on late health effects following radiotherapy, we sought to demonstrate a deep learning approach to predict 3D whole-body dose maps for pediatric patients. Materials/

Methods: We used a database of Monte Carlo-derived whole-body dose maps which were previously reconstructed for 4716 pediatric patients in the National Wilms Tumor Study late effects cohort. The patients in this cohort tended to receive standardized treatments with the field borders determined by anatomical structures. To train and test our model we identified subsets of patients who received a 6 MV AP-PA beam configuration to the abdomen (with or without kidney block), left flank, right flank, or chest (90% patients used for training, 10% for testing). A customized neural network was trained separately for each field configuration with the dose map being normalized by the prescribed dose. The input to our model was: (1) a CT image of a whole-body phantom matched to the patient based on height and weight; (2) a RT structure file containing organ contours of interest; (3) the normalized dose map; and (4) the patient’s height, weight, and gender. Next, we extracted multi-level and multi-scale feature maps, the predicted dose map file, from the customized deep neural network final layers. Predicted doses were compared voxel-to-voxel with original doses using different statistical metrics. Dose-volume histograms were made for each organ at risk to visualize differences between actual and predicted doses.
Results: The absolute dose map for a patient was calculated by multiplying the predicted normalized dose map by the patient’s radiotherapy prescribed dose. The dose maps for patients receiving radiation to more than one location could be predicted separately and added together to get the total from the entire plan. Organ dose was calculated by applying organ contour masks to the predicted dose map. Our quantitative analysis of the test dataset yielded a structural similarity index of 0.89±0.02 (mean ± SD) Confidence Interval (CI): [0.86, 0.90], mean absolute error of 2.57±0.49 Gy (CI: [2.17, 3.40]), and root mean square error of 9.28±1.97 Gy (CI: [7.54, 12.40]).
Conclusion:

Our results suggest that the algorithm is functional, although further refinement is necessary as error rates for some patients exceed expectations. We are also exploring extending this work to multi-class-based dose prediction considering the additional photon treatment energies (e.g., 4 MV, 10 MV, 16 MV, and Co-60).